Claims about continual learning are easy to make and hard to verify. Almost any model can look good if you let it keep a copy of the old data and quietly re-train on it, or if you only test on the most recent task. So we set ourselves stricter rules: public benchmarks, strong dense baselines, and no replay buffer at all. The model sees each domain once, in sequence, and then has to remember it.

What we measured

Continual-learning results are usually reported with two numbers, and you need both:

  • Retention is how much of what the model learned on earlier tasks it still knows after learning later ones.
  • Forgetting is how far performance on an earlier task dropped from its best. Lower is better.

A model can cheat one of these by sacrificing the other: refuse to learn anything new and you'll never forget. That's why we also report accuracy alongside, so you can see Voltaic isn't buying stability by standing still.

Results

CORe50-NI: forgetting falls to 0.04%

CORe50-NI presents the same object classes under new conditions, new backgrounds, lighting and poses, one batch at a time. It's a good stand-in for the real world, where things you already know keep turning up looking slightly different. A standard model forgot 1.20%. Voltaic with HULL forgot 0.04%, a thirtyfold reduction. Average accuracy didn't fall to pay for it; it rose, from 95.01% to 95.23%.

Split CIFAR-100: 2.6× more retained

Split CIFAR-100 divides 100 image classes into a sequence of tasks the model must learn one after another. Voltaic retained 15.74%, against 6.14% (Qwen 7B) and 5.91% (Mistral 7B) for two dense 7B baselines: 2.6× more knowledge kept.

Split ImageNet-1k: 4.2× more retained

The same protocol at ImageNet scale, with a thousand classes, is far harder. Voltaic retained 20.87%, against 4.93% (Qwen 7B) and 4.90% (Mistral 7B) for the dense 7B baselines: 4.2× more. Notably, the gap widens as the problem gets larger, which is the opposite of what you'd expect if Voltaic's advantage were a small-benchmark artefact.

Protocol

We want these numbers to be reproducible, so here is exactly how they were produced:

  • Language and CORe50 tests: CLBench v10, no-replay sequential protocol, single A100, run on 30 June 2026 with seed 0.
  • Vision tests: 8× A100, run on 14 June 2026. Images are encoded with frozen, pretrained DINOv2 features; continual learning happens in Voltaic's spiking core on top of them.
  • Baselines: two dense 7B models, Qwen 7B and Mistral 7B, run under the same protocol.
  • Seeds: these are single-seed results. Multi-seed results will be published alongside the paper.

No replay buffer means no stored copies of earlier tasks' data. That's the hardest setting and the one that matters for privacy: a model that needs to keep your old data around to avoid forgetting it isn't really learning continually.

What HULL adds

The Accumulative Hull Neuron Model (HULL) keeps a geometric "stability hull" around skills the model has mastered. Think of each mastered skill as occupying a region of the network's weight space. HULL tracks the boundary of that region and steers new learning so it happens outside it. Old skills stay protected, and because the network keeps refining the representation around them, they often sharpen. Across tasks we measure an average drift of 0.71×: below one means old performance tends to improve, not decay, as new tasks arrive.

HULL works together with Hierarchical Synaptic Consolidation, which decides at the level of individual synapses what deserves to be hardened. We explain that in Living Memory.

What this means outside the lab

Benchmarks are a proxy. The real test is whether a model can be taught something on Tuesday and still know Monday's lesson. In practice these results mean you can keep teaching Voltaic, new products, new people, new processes, without scheduling a retraining run and without worrying that last quarter's knowledge is quietly eroding. And it matters even more for robots: a machine that learns a new task in the field must not unlearn how to walk. That's exactly the property CORe50-NI measures.

Read how the architecture works in Voltaic: a mind that learns while you use it, or see the full figures on the Voltaic page.